Capability ≠ trust
A capable model is not trusted by default. Each agent works inside explicit permissions, resources and boundaries set for the task.
Human Minority · AI agents under human authority
Human MinorityMore agents. Same veto.
Human Minority is a governing system for AI agents that stays subordinate to a human. The human sets the goal and keeps the final approve, reject and veto. The system organises the agents' work, holds them to their permissions and budgets, and checks their results independently.
AI agents already produce work faster than a person can check it. Adding more agents makes that gap wider. Human Minority is built to close it without taking the last word away from the human.
Claim is not proof.
Early development · not yet a released product
Three principles
Every part of the system follows from three separations. An agent's strength, its own report and its track record are useful signals. None of them is, by itself, permission to act or proof that the work is right.
Claim ≠ proof
"Tests passed" or "ready to merge" from the agent that did the work is a claim. Results are checked independently against the exact output.
Reputation ≠ authority
A good history can inform who gets the next task. It never replaces the human's authority to approve, reject or stop.
How it is meant to work
This section describes the intended product. Most of it is in development and not yet public; see Status for what exists today.
The human sets the goal and the limits. What should be done, by whom it may be done, and with which permissions and budget.
The system assigns and supervises the agents. It selects agents for the task, organises their work and enforces their permissions and resources while they run.
Evidence is collected and checked independently. Results are bound to the exact output and checked by a verifier that is not the agent that produced them.
Verified history is kept. What happened is recorded so it can be reviewed later and used to improve future assignments.
The human decides. Instead of a flood of agent output, the human gets a structured decision package and keeps the final word.
What the human receives
The goal is more human capability without a matching growth in review load. Each decision package answers four questions.
What was done
The exact work the agents produced, tied to the task it was meant to complete.
What is backed by evidence
Which results were independently checked, and which are still only claims.
What risks remain
Open questions, failed or missing checks, and anything the system could not establish.
Which decisions are needed
Exactly what the human has to approve, reject or decide next.
Software development is the first application, because its results can be checked precisely. The approach itself is not limited to software.
Status today
Human Minority is in early development. It is not yet available as a product, and no release date is announced.
Available now: Human Minority Verifier
The public verification component. It binds an agent's claim, the exact candidate bytes, the required checks and the verifier results, and decides from those bindings rather than from the agent's own report. Source-available under PolyForm Perimeter 1.0.1; not OSI open source.
In development
- assigning and supervising agents from several providers;
- enforcing permissions, resources and budgets during a run;
- the human decision package and the approve / reject / veto flow;
- verified history and improvement of future assignments.
These parts are not public and are not covered by any availability or security claim on this page.